A convolutional neural network (CNN) was found to skillfully classify potentially severe convection of a future climate based on learned thermodynamic and kinematic thunderstorm features. The CNN was trained to classify strongly rotating storms from a current climate, then evaluated against storms from a future climate (end of 21^st century), and found to perform with skill and comparatively in both climates. Strongly rotating storms were of interest because they are more likely to be supercells, a thunderstorm type that has a greater likelihood of producing tornadoes and large hail, which cause billions of losses and dozens of fatalities every year. Despite training with labels derived from a threshold value of a severe thunderstorm diagnostic (updraft helicity), the CNN learned physical characteristics of organized convection and environments that are not captured by the diagnostic heuristic. Interpretability techniques revealed that strongly rotating storms are associated with rotation signatures and thunderstorm updrafts penetrating comparatively drier vertical mid-levels. Results show that simple heuristics can yield skillful results with CNNs and can be used to generate labeled data for supervised learning frameworks. Most importantly, results from this study show that deep learning is capable of generalizing to future climate extremes and can exhibit out-of-sample robustness with proper hyperparameter tuning. As the climate continues to change, and machine learning techniques continue to proliferate in the physical sciences, it is important to ensure that techniques perform skillfully with unseen outliers and climate signals. This study offers evidence that this objective is possible and based on physical signals.